1#include "rs_core.rsh" 2#include "rs_f16_util.h" 3 4extern float2 __attribute__((overloadable)) convert_float2(int2 c); 5extern float3 __attribute__((overloadable)) convert_float3(int3 c); 6extern float4 __attribute__((overloadable)) convert_float4(int4 c); 7 8extern int2 __attribute__((overloadable)) convert_int2(float2 c); 9extern int3 __attribute__((overloadable)) convert_int3(float3 c); 10extern int4 __attribute__((overloadable)) convert_int4(float4 c); 11 12 13extern float __attribute__((overloadable)) fmin(float v, float v2); 14extern float2 __attribute__((overloadable)) fmin(float2 v, float v2); 15extern float3 __attribute__((overloadable)) fmin(float3 v, float v2); 16extern float4 __attribute__((overloadable)) fmin(float4 v, float v2); 17 18extern float __attribute__((overloadable)) fmax(float v, float v2); 19extern float2 __attribute__((overloadable)) fmax(float2 v, float v2); 20extern float3 __attribute__((overloadable)) fmax(float3 v, float v2); 21extern float4 __attribute__((overloadable)) fmax(float4 v, float v2); 22 23// Float ops, 6.11.2 24 25#define FN_FUNC_FN(fnc) \ 26extern float2 __attribute__((overloadable)) fnc(float2 v) { \ 27 float2 r; \ 28 r.x = fnc(v.x); \ 29 r.y = fnc(v.y); \ 30 return r; \ 31} \ 32extern float3 __attribute__((overloadable)) fnc(float3 v) { \ 33 float3 r; \ 34 r.x = fnc(v.x); \ 35 r.y = fnc(v.y); \ 36 r.z = fnc(v.z); \ 37 return r; \ 38} \ 39extern float4 __attribute__((overloadable)) fnc(float4 v) { \ 40 float4 r; \ 41 r.x = fnc(v.x); \ 42 r.y = fnc(v.y); \ 43 r.z = fnc(v.z); \ 44 r.w = fnc(v.w); \ 45 return r; \ 46} 47 48#define IN_FUNC_FN(fnc) \ 49extern int2 __attribute__((overloadable)) fnc(float2 v) { \ 50 int2 r; \ 51 r.x = fnc(v.x); \ 52 r.y = fnc(v.y); \ 53 return r; \ 54} \ 55extern int3 __attribute__((overloadable)) fnc(float3 v) { \ 56 int3 r; \ 57 r.x = fnc(v.x); \ 58 r.y = fnc(v.y); \ 59 r.z = fnc(v.z); \ 60 return r; \ 61} \ 62extern int4 __attribute__((overloadable)) fnc(float4 v) { \ 63 int4 r; \ 64 r.x = fnc(v.x); \ 65 r.y = fnc(v.y); \ 66 r.z = fnc(v.z); \ 67 r.w = fnc(v.w); \ 68 return r; \ 69} 70 71#define FN_FUNC_FN_FN(fnc) \ 72extern float2 __attribute__((overloadable)) fnc(float2 v1, float2 v2) { \ 73 float2 r; \ 74 r.x = fnc(v1.x, v2.x); \ 75 r.y = fnc(v1.y, v2.y); \ 76 return r; \ 77} \ 78extern float3 __attribute__((overloadable)) fnc(float3 v1, float3 v2) { \ 79 float3 r; \ 80 r.x = fnc(v1.x, v2.x); \ 81 r.y = fnc(v1.y, v2.y); \ 82 r.z = fnc(v1.z, v2.z); \ 83 return r; \ 84} \ 85extern float4 __attribute__((overloadable)) fnc(float4 v1, float4 v2) { \ 86 float4 r; \ 87 r.x = fnc(v1.x, v2.x); \ 88 r.y = fnc(v1.y, v2.y); \ 89 r.z = fnc(v1.z, v2.z); \ 90 r.w = fnc(v1.w, v2.w); \ 91 return r; \ 92} 93 94#define FN_FUNC_FN_F(fnc) \ 95extern float2 __attribute__((overloadable)) fnc(float2 v1, float v2) { \ 96 float2 r; \ 97 r.x = fnc(v1.x, v2); \ 98 r.y = fnc(v1.y, v2); \ 99 return r; \ 100} \ 101extern float3 __attribute__((overloadable)) fnc(float3 v1, float v2) { \ 102 float3 r; \ 103 r.x = fnc(v1.x, v2); \ 104 r.y = fnc(v1.y, v2); \ 105 r.z = fnc(v1.z, v2); \ 106 return r; \ 107} \ 108extern float4 __attribute__((overloadable)) fnc(float4 v1, float v2) { \ 109 float4 r; \ 110 r.x = fnc(v1.x, v2); \ 111 r.y = fnc(v1.y, v2); \ 112 r.z = fnc(v1.z, v2); \ 113 r.w = fnc(v1.w, v2); \ 114 return r; \ 115} 116 117#define FN_FUNC_FN_IN(fnc) \ 118extern float2 __attribute__((overloadable)) fnc(float2 v1, int2 v2) { \ 119 float2 r; \ 120 r.x = fnc(v1.x, v2.x); \ 121 r.y = fnc(v1.y, v2.y); \ 122 return r; \ 123} \ 124extern float3 __attribute__((overloadable)) fnc(float3 v1, int3 v2) { \ 125 float3 r; \ 126 r.x = fnc(v1.x, v2.x); \ 127 r.y = fnc(v1.y, v2.y); \ 128 r.z = fnc(v1.z, v2.z); \ 129 return r; \ 130} \ 131extern float4 __attribute__((overloadable)) fnc(float4 v1, int4 v2) { \ 132 float4 r; \ 133 r.x = fnc(v1.x, v2.x); \ 134 r.y = fnc(v1.y, v2.y); \ 135 r.z = fnc(v1.z, v2.z); \ 136 r.w = fnc(v1.w, v2.w); \ 137 return r; \ 138} 139 140#define FN_FUNC_FN_I(fnc) \ 141extern float2 __attribute__((overloadable)) fnc(float2 v1, int v2) { \ 142 float2 r; \ 143 r.x = fnc(v1.x, v2); \ 144 r.y = fnc(v1.y, v2); \ 145 return r; \ 146} \ 147extern float3 __attribute__((overloadable)) fnc(float3 v1, int v2) { \ 148 float3 r; \ 149 r.x = fnc(v1.x, v2); \ 150 r.y = fnc(v1.y, v2); \ 151 r.z = fnc(v1.z, v2); \ 152 return r; \ 153} \ 154extern float4 __attribute__((overloadable)) fnc(float4 v1, int v2) { \ 155 float4 r; \ 156 r.x = fnc(v1.x, v2); \ 157 r.y = fnc(v1.y, v2); \ 158 r.z = fnc(v1.z, v2); \ 159 r.w = fnc(v1.w, v2); \ 160 return r; \ 161} 162 163#define FN_FUNC_FN_PFN(fnc) \ 164extern float2 __attribute__((overloadable)) \ 165 fnc(float2 v1, float2 *v2) { \ 166 float2 r; \ 167 float t[2]; \ 168 r.x = fnc(v1.x, &t[0]); \ 169 r.y = fnc(v1.y, &t[1]); \ 170 v2->x = t[0]; \ 171 v2->y = t[1]; \ 172 return r; \ 173} \ 174extern float3 __attribute__((overloadable)) \ 175 fnc(float3 v1, float3 *v2) { \ 176 float3 r; \ 177 float t[3]; \ 178 r.x = fnc(v1.x, &t[0]); \ 179 r.y = fnc(v1.y, &t[1]); \ 180 r.z = fnc(v1.z, &t[2]); \ 181 v2->x = t[0]; \ 182 v2->y = t[1]; \ 183 v2->z = t[2]; \ 184 return r; \ 185} \ 186extern float4 __attribute__((overloadable)) \ 187 fnc(float4 v1, float4 *v2) { \ 188 float4 r; \ 189 float t[4]; \ 190 r.x = fnc(v1.x, &t[0]); \ 191 r.y = fnc(v1.y, &t[1]); \ 192 r.z = fnc(v1.z, &t[2]); \ 193 r.w = fnc(v1.w, &t[3]); \ 194 v2->x = t[0]; \ 195 v2->y = t[1]; \ 196 v2->z = t[2]; \ 197 v2->w = t[3]; \ 198 return r; \ 199} 200 201#define FN_FUNC_FN_PIN(fnc) \ 202extern float2 __attribute__((overloadable)) fnc(float2 v1, int2 *v2) { \ 203 float2 r; \ 204 int t[2]; \ 205 r.x = fnc(v1.x, &t[0]); \ 206 r.y = fnc(v1.y, &t[1]); \ 207 v2->x = t[0]; \ 208 v2->y = t[1]; \ 209 return r; \ 210} \ 211extern float3 __attribute__((overloadable)) fnc(float3 v1, int3 *v2) { \ 212 float3 r; \ 213 int t[3]; \ 214 r.x = fnc(v1.x, &t[0]); \ 215 r.y = fnc(v1.y, &t[1]); \ 216 r.z = fnc(v1.z, &t[2]); \ 217 v2->x = t[0]; \ 218 v2->y = t[1]; \ 219 v2->z = t[2]; \ 220 return r; \ 221} \ 222extern float4 __attribute__((overloadable)) fnc(float4 v1, int4 *v2) { \ 223 float4 r; \ 224 int t[4]; \ 225 r.x = fnc(v1.x, &t[0]); \ 226 r.y = fnc(v1.y, &t[1]); \ 227 r.z = fnc(v1.z, &t[2]); \ 228 r.w = fnc(v1.w, &t[3]); \ 229 v2->x = t[0]; \ 230 v2->y = t[1]; \ 231 v2->z = t[2]; \ 232 v2->w = t[3]; \ 233 return r; \ 234} 235 236#define FN_FUNC_FN_FN_FN(fnc) \ 237extern float2 __attribute__((overloadable)) \ 238 fnc(float2 v1, float2 v2, float2 v3) { \ 239 float2 r; \ 240 r.x = fnc(v1.x, v2.x, v3.x); \ 241 r.y = fnc(v1.y, v2.y, v3.y); \ 242 return r; \ 243} \ 244extern float3 __attribute__((overloadable)) \ 245 fnc(float3 v1, float3 v2, float3 v3) { \ 246 float3 r; \ 247 r.x = fnc(v1.x, v2.x, v3.x); \ 248 r.y = fnc(v1.y, v2.y, v3.y); \ 249 r.z = fnc(v1.z, v2.z, v3.z); \ 250 return r; \ 251} \ 252extern float4 __attribute__((overloadable)) \ 253 fnc(float4 v1, float4 v2, float4 v3) { \ 254 float4 r; \ 255 r.x = fnc(v1.x, v2.x, v3.x); \ 256 r.y = fnc(v1.y, v2.y, v3.y); \ 257 r.z = fnc(v1.z, v2.z, v3.z); \ 258 r.w = fnc(v1.w, v2.w, v3.w); \ 259 return r; \ 260} 261 262#define FN_FUNC_FN_FN_PIN(fnc) \ 263extern float2 __attribute__((overloadable)) \ 264 fnc(float2 v1, float2 v2, int2 *v3) { \ 265 float2 r; \ 266 int t[2]; \ 267 r.x = fnc(v1.x, v2.x, &t[0]); \ 268 r.y = fnc(v1.y, v2.y, &t[1]); \ 269 v3->x = t[0]; \ 270 v3->y = t[1]; \ 271 return r; \ 272} \ 273extern float3 __attribute__((overloadable)) \ 274 fnc(float3 v1, float3 v2, int3 *v3) { \ 275 float3 r; \ 276 int t[3]; \ 277 r.x = fnc(v1.x, v2.x, &t[0]); \ 278 r.y = fnc(v1.y, v2.y, &t[1]); \ 279 r.z = fnc(v1.z, v2.z, &t[2]); \ 280 v3->x = t[0]; \ 281 v3->y = t[1]; \ 282 v3->z = t[2]; \ 283 return r; \ 284} \ 285extern float4 __attribute__((overloadable)) \ 286 fnc(float4 v1, float4 v2, int4 *v3) { \ 287 float4 r; \ 288 int t[4]; \ 289 r.x = fnc(v1.x, v2.x, &t[0]); \ 290 r.y = fnc(v1.y, v2.y, &t[1]); \ 291 r.z = fnc(v1.z, v2.z, &t[2]); \ 292 r.w = fnc(v1.w, v2.w, &t[3]); \ 293 v3->x = t[0]; \ 294 v3->y = t[1]; \ 295 v3->z = t[2]; \ 296 v3->w = t[3]; \ 297 return r; \ 298} 299 300static const int iposinf = 0x7f800000; 301static const int ineginf = 0xff800000; 302 303static const float posinf() { 304 float f = *((float*)&iposinf); 305 return f; 306} 307 308static const float neginf() { 309 float f = *((float*)&ineginf); 310 return f; 311} 312 313static bool isinf(float f) { 314 int i = *((int*)(void*)&f); 315 return (i == iposinf) || (i == ineginf); 316} 317 318static bool isnan(float f) { 319 int i = *((int*)(void*)&f); 320 return (((i & 0x7f800000) == 0x7f800000) && (i & 0x007fffff)); 321} 322 323static bool isposzero(float f) { 324 int i = *((int*)(void*)&f); 325 return (i == 0x00000000); 326} 327 328static bool isnegzero(float f) { 329 int i = *((int*)(void*)&f); 330 return (i == 0x80000000); 331} 332 333static bool iszero(float f) { 334 return isposzero(f) || isnegzero(f); 335} 336 337 338extern float __attribute__((overloadable)) SC_acosf(float); 339float __attribute__((overloadable)) acos(float v) { 340 return SC_acosf(v); 341} 342FN_FUNC_FN(acos) 343 344extern float __attribute__((overloadable)) SC_acoshf(float); 345float __attribute__((overloadable)) acosh(float v) { 346 return SC_acoshf(v); 347} 348FN_FUNC_FN(acosh) 349 350 351extern float __attribute__((overloadable)) acospi(float v) { 352 return acos(v) / M_PI; 353} 354FN_FUNC_FN(acospi) 355 356extern float __attribute__((overloadable)) SC_asinf(float); 357float __attribute__((overloadable)) asin(float v) { 358 return SC_asinf(v); 359} 360FN_FUNC_FN(asin) 361 362extern float __attribute__((overloadable)) SC_asinhf(float); 363float __attribute__((overloadable)) asinh(float v) { 364 return SC_asinhf(v); 365} 366FN_FUNC_FN(asinh) 367 368extern float __attribute__((overloadable)) asinpi(float v) { 369 return asin(v) / M_PI; 370} 371FN_FUNC_FN(asinpi) 372 373extern float __attribute__((overloadable)) SC_atanf(float); 374float __attribute__((overloadable)) atan(float v) { 375 return SC_atanf(v); 376} 377FN_FUNC_FN(atan) 378 379extern float __attribute__((overloadable)) SC_atan2f(float, float); 380float __attribute__((overloadable)) atan2(float v1, float v2) { 381 return SC_atan2f(v1, v2); 382} 383FN_FUNC_FN_FN(atan2) 384 385extern float __attribute__((overloadable)) SC_atanhf(float); 386float __attribute__((overloadable)) atanh(float v) { 387 return SC_atanhf(v); 388} 389FN_FUNC_FN(atanh) 390 391extern float __attribute__((overloadable)) atanpi(float v) { 392 return atan(v) / M_PI; 393} 394FN_FUNC_FN(atanpi) 395 396 397extern float __attribute__((overloadable)) atan2pi(float y, float x) { 398 return atan2(y, x) / M_PI; 399} 400FN_FUNC_FN_FN(atan2pi) 401 402extern float __attribute__((overloadable)) SC_cbrtf(float); 403float __attribute__((overloadable)) cbrt(float v) { 404 return SC_cbrtf(v); 405} 406FN_FUNC_FN(cbrt) 407 408extern float __attribute__((overloadable)) SC_ceilf(float); 409float __attribute__((overloadable)) ceil(float v) { 410 return SC_ceilf(v); 411} 412FN_FUNC_FN(ceil) 413 414extern float __attribute__((overloadable)) SC_copysignf(float, float); 415float __attribute__((overloadable)) copysign(float v1, float v2) { 416 return SC_copysignf(v1, v2); 417} 418FN_FUNC_FN_FN(copysign) 419 420extern float __attribute__((overloadable)) SC_cosf(float); 421float __attribute__((overloadable)) cos(float v) { 422 return SC_cosf(v); 423} 424FN_FUNC_FN(cos) 425 426extern float __attribute__((overloadable)) SC_coshf(float); 427float __attribute__((overloadable)) cosh(float v) { 428 return SC_coshf(v); 429} 430FN_FUNC_FN(cosh) 431 432extern float __attribute__((overloadable)) cospi(float v) { 433 return cos(v * M_PI); 434} 435FN_FUNC_FN(cospi) 436 437extern float __attribute__((overloadable)) SC_erfcf(float); 438float __attribute__((overloadable)) erfc(float v) { 439 return SC_erfcf(v); 440} 441FN_FUNC_FN(erfc) 442 443extern float __attribute__((overloadable)) SC_erff(float); 444float __attribute__((overloadable)) erf(float v) { 445 return SC_erff(v); 446} 447FN_FUNC_FN(erf) 448 449extern float __attribute__((overloadable)) SC_expf(float); 450float __attribute__((overloadable)) exp(float v) { 451 return SC_expf(v); 452} 453FN_FUNC_FN(exp) 454 455extern float __attribute__((overloadable)) SC_exp2f(float); 456float __attribute__((overloadable)) exp2(float v) { 457 return SC_exp2f(v); 458} 459FN_FUNC_FN(exp2) 460 461extern float __attribute__((overloadable)) pow(float, float); 462 463extern float __attribute__((overloadable)) exp10(float v) { 464 return exp2(v * 3.321928095f); 465} 466FN_FUNC_FN(exp10) 467 468extern float __attribute__((overloadable)) SC_expm1f(float); 469float __attribute__((overloadable)) expm1(float v) { 470 return SC_expm1f(v); 471} 472FN_FUNC_FN(expm1) 473 474extern float __attribute__((overloadable)) fabs(float v) { 475 int i = *((int*)(void*)&v) & 0x7fffffff; 476 return *((float*)(void*)&i); 477} 478FN_FUNC_FN(fabs) 479 480extern float __attribute__((overloadable)) SC_fdimf(float, float); 481float __attribute__((overloadable)) fdim(float v1, float v2) { 482 return SC_fdimf(v1, v2); 483} 484FN_FUNC_FN_FN(fdim) 485 486extern float __attribute__((overloadable)) SC_floorf(float); 487float __attribute__((overloadable)) floor(float v) { 488 return SC_floorf(v); 489} 490FN_FUNC_FN(floor) 491 492extern float __attribute__((overloadable)) SC_fmaf(float, float, float); 493float __attribute__((overloadable)) fma(float v1, float v2, float v3) { 494 return SC_fmaf(v1, v2, v3); 495} 496FN_FUNC_FN_FN_FN(fma) 497 498extern float __attribute__((overloadable)) SC_fminf(float, float); 499 500extern float __attribute__((overloadable)) SC_fmodf(float, float); 501float __attribute__((overloadable)) fmod(float v1, float v2) { 502 return SC_fmodf(v1, v2); 503} 504FN_FUNC_FN_FN(fmod) 505 506extern float __attribute__((overloadable)) fract(float v, float *iptr) { 507 int i = (int)floor(v); 508 if (iptr) { 509 iptr[0] = i; 510 } 511 return fmin(v - i, 0x1.fffffep-1f); 512} 513FN_FUNC_FN_PFN(fract) 514 515extern float __attribute__((const, overloadable)) fract(float v) { 516 float unused; 517 return fract(v, &unused); 518} 519FN_FUNC_FN(fract) 520 521extern float __attribute__((overloadable)) SC_frexpf(float, int *); 522float __attribute__((overloadable)) frexp(float v1, int* v2) { 523 return SC_frexpf(v1, v2); 524} 525FN_FUNC_FN_PIN(frexp) 526 527extern float __attribute__((overloadable)) SC_hypotf(float, float); 528float __attribute__((overloadable)) hypot(float v1, float v2) { 529 return SC_hypotf(v1, v2); 530} 531FN_FUNC_FN_FN(hypot) 532 533extern int __attribute__((overloadable)) SC_ilogbf(float); 534int __attribute__((overloadable)) ilogb(float v) { 535 return SC_ilogbf(v); 536} 537IN_FUNC_FN(ilogb) 538 539extern float __attribute__((overloadable)) SC_ldexpf(float, int); 540float __attribute__((overloadable)) ldexp(float v1, int v2) { 541 return SC_ldexpf(v1, v2); 542} 543FN_FUNC_FN_IN(ldexp) 544FN_FUNC_FN_I(ldexp) 545 546extern float __attribute__((overloadable)) SC_lgammaf(float); 547float __attribute__((overloadable)) lgamma(float v) { 548 return SC_lgammaf(v); 549} 550FN_FUNC_FN(lgamma) 551extern float __attribute__((overloadable)) SC_lgammaf_r(float, int*); 552float __attribute__((overloadable)) lgamma(float v, int* ptr) { 553 return SC_lgammaf_r(v, ptr); 554} 555FN_FUNC_FN_PIN(lgamma) 556 557extern float __attribute__((overloadable)) SC_logf(float); 558float __attribute__((overloadable)) log(float v) { 559 return SC_logf(v); 560} 561FN_FUNC_FN(log) 562 563extern float __attribute__((overloadable)) SC_log10f(float); 564float __attribute__((overloadable)) log10(float v) { 565 return SC_log10f(v); 566} 567FN_FUNC_FN(log10) 568 569 570extern float __attribute__((overloadable)) log2(float v) { 571 return log10(v) * 3.321928095f; 572} 573FN_FUNC_FN(log2) 574 575extern float __attribute__((overloadable)) SC_log1pf(float); 576float __attribute__((overloadable)) log1p(float v) { 577 return SC_log1pf(v); 578} 579FN_FUNC_FN(log1p) 580 581extern float __attribute__((overloadable)) SC_logbf(float); 582float __attribute__((overloadable)) logb(float v) { 583 return SC_logbf(v); 584} 585FN_FUNC_FN(logb) 586 587extern float __attribute__((overloadable)) mad(float a, float b, float c) { 588 return a * b + c; 589} 590extern float2 __attribute__((overloadable)) mad(float2 a, float2 b, float2 c) { 591 return a * b + c; 592} 593extern float3 __attribute__((overloadable)) mad(float3 a, float3 b, float3 c) { 594 return a * b + c; 595} 596extern float4 __attribute__((overloadable)) mad(float4 a, float4 b, float4 c) { 597 return a * b + c; 598} 599 600extern float __attribute__((overloadable)) SC_modff(float, float *); 601float __attribute__((overloadable)) modf(float v1, float *v2) { 602 return SC_modff(v1, v2); 603} 604FN_FUNC_FN_PFN(modf); 605 606extern float __attribute__((overloadable)) nan(uint v) { 607 float f[1]; 608 uint32_t *ip = (uint32_t *)f; 609 *ip = v | 0x7fc00000; 610 return f[0]; 611} 612 613extern float __attribute__((overloadable)) SC_nextafterf(float, float); 614float __attribute__((overloadable)) nextafter(float v1, float v2) { 615 return SC_nextafterf(v1, v2); 616} 617FN_FUNC_FN_FN(nextafter) 618 619// This function must be defined here if we're compiling with debug info 620// (libclcore_g.bc), because we need a C source to get debug information. 621// Otherwise the implementation can be found in IR. 622#if defined(RS_G_RUNTIME) 623extern float __attribute__((overloadable)) SC_powf(float, float); 624float __attribute__((overloadable)) pow(float v1, float v2) { 625 return SC_powf(v1, v2); 626} 627#endif // defined(RS_G_RUNTIME) 628FN_FUNC_FN_FN(pow) 629 630extern float __attribute__((overloadable)) pown(float v, int p) { 631 /* The mantissa of a float has fewer bits than an int (24 effective vs. 31). 632 * For very large ints, we'll lose whether the exponent is even or odd, making 633 * the selection of a correct sign incorrect. We correct this. Use copysign 634 * to handle the negative zero case. 635 */ 636 float sign = (p & 0x1) ? copysign(1.f, v) : 1.f; 637 float f = pow(v, (float)p); 638 return copysign(f, sign); 639} 640FN_FUNC_FN_IN(pown) 641 642extern float __attribute__((overloadable)) powr(float v, float p) { 643 return pow(v, p); 644} 645extern float2 __attribute__((overloadable)) powr(float2 v, float2 p) { 646 return pow(v, p); 647} 648extern float3 __attribute__((overloadable)) powr(float3 v, float3 p) { 649 return pow(v, p); 650} 651extern float4 __attribute__((overloadable)) powr(float4 v, float4 p) { 652 return pow(v, p); 653} 654 655extern float __attribute__((overloadable)) SC_remainderf(float, float); 656float __attribute__((overloadable)) remainder(float v1, float v2) { 657 return SC_remainderf(v1, v2); 658} 659FN_FUNC_FN_FN(remainder) 660 661extern float __attribute__((overloadable)) SC_remquof(float, float, int *); 662float __attribute__((overloadable)) remquo(float v1, float v2, int *v3) { 663 return SC_remquof(v1, v2, v3); 664} 665FN_FUNC_FN_FN_PIN(remquo) 666 667extern float __attribute__((overloadable)) SC_rintf(float); 668float __attribute__((overloadable)) rint(float v) { 669 return SC_rintf(v); 670} 671FN_FUNC_FN(rint) 672 673extern float __attribute__((overloadable)) rootn(float v, int r) { 674 if (r == 0) { 675 return posinf(); 676 } 677 678 if (iszero(v)) { 679 if (r < 0) { 680 if (r & 1) { 681 return copysign(posinf(), v); 682 } else { 683 return posinf(); 684 } 685 } else { 686 if (r & 1) { 687 return copysign(0.f, v); 688 } else { 689 return 0.f; 690 } 691 } 692 } 693 694 if (!isinf(v) && !isnan(v) && (v < 0.f)) { 695 if (r & 1) { 696 return (-1.f * pow(-1.f * v, 1.f / r)); 697 } else { 698 return nan(0); 699 } 700 } 701 702 return pow(v, 1.f / r); 703} 704FN_FUNC_FN_IN(rootn); 705 706extern float __attribute__((overloadable)) SC_roundf(float); 707float __attribute__((overloadable)) round(float v) { 708 return SC_roundf(v); 709} 710FN_FUNC_FN(round) 711 712extern float __attribute__((overloadable)) SC_randf2(float, float); 713float __attribute__((overloadable)) rsRand(float min, float max) { 714 return SC_randf2(min, max); 715} 716 717 718extern float __attribute__((overloadable)) rsqrt(float v) { 719 return 1.f / sqrt(v); 720} 721 722#if !defined(ARCH_X86_HAVE_SSSE3) || defined(RS_DEBUG_RUNTIME) || defined(RS_G_RUNTIME) 723// These functions must be defined here if we are not using the SSE 724// implementation, which includes when we are built as part of the 725// debug runtime (libclcore_debug.bc) or compiling with debug info. 726#if defined(RS_G_RUNTIME) 727extern float __attribute__((overloadable)) SC_sqrtf(float); 728float __attribute__((overloadable)) sqrt(float v) { 729 return SC_sqrtf(v); 730} 731#endif // defined(RS_G_RUNTIME) 732 733FN_FUNC_FN(sqrt) 734#else 735extern float2 __attribute__((overloadable)) sqrt(float2); 736extern float3 __attribute__((overloadable)) sqrt(float3); 737extern float4 __attribute__((overloadable)) sqrt(float4); 738#endif // !defined(ARCH_X86_HAVE_SSSE3) || defined(RS_DEBUG_RUNTIME) || defined(RS_G_RUNTIME) 739 740FN_FUNC_FN(rsqrt) 741 742extern float __attribute__((overloadable)) SC_sinf(float); 743float __attribute__((overloadable)) sin(float v) { 744 return SC_sinf(v); 745} 746FN_FUNC_FN(sin) 747 748extern float __attribute__((overloadable)) sincos(float v, float *cosptr) { 749 *cosptr = cos(v); 750 return sin(v); 751} 752extern float2 __attribute__((overloadable)) sincos(float2 v, float2 *cosptr) { 753 *cosptr = cos(v); 754 return sin(v); 755} 756extern float3 __attribute__((overloadable)) sincos(float3 v, float3 *cosptr) { 757 *cosptr = cos(v); 758 return sin(v); 759} 760extern float4 __attribute__((overloadable)) sincos(float4 v, float4 *cosptr) { 761 *cosptr = cos(v); 762 return sin(v); 763} 764 765extern float __attribute__((overloadable)) SC_sinhf(float); 766float __attribute__((overloadable)) sinh(float v) { 767 return SC_sinhf(v); 768} 769FN_FUNC_FN(sinh) 770 771extern float __attribute__((overloadable)) sinpi(float v) { 772 return sin(v * M_PI); 773} 774FN_FUNC_FN(sinpi) 775 776extern float __attribute__((overloadable)) SC_tanf(float); 777float __attribute__((overloadable)) tan(float v) { 778 return SC_tanf(v); 779} 780FN_FUNC_FN(tan) 781 782extern float __attribute__((overloadable)) SC_tanhf(float); 783float __attribute__((overloadable)) tanh(float v) { 784 return SC_tanhf(v); 785} 786FN_FUNC_FN(tanh) 787 788extern float __attribute__((overloadable)) tanpi(float v) { 789 return tan(v * M_PI); 790} 791FN_FUNC_FN(tanpi) 792 793 794extern float __attribute__((overloadable)) SC_tgammaf(float); 795float __attribute__((overloadable)) tgamma(float v) { 796 return SC_tgammaf(v); 797} 798FN_FUNC_FN(tgamma) 799 800extern float __attribute__((overloadable)) SC_truncf(float); 801float __attribute__((overloadable)) trunc(float v) { 802 return SC_truncf(v); 803} 804FN_FUNC_FN(trunc) 805 806// Int ops (partial), 6.11.3 807 808#define XN_FUNC_YN(typeout, fnc, typein) \ 809extern typeout __attribute__((overloadable)) fnc(typein); \ 810extern typeout##2 __attribute__((overloadable)) fnc(typein##2 v) { \ 811 typeout##2 r; \ 812 r.x = fnc(v.x); \ 813 r.y = fnc(v.y); \ 814 return r; \ 815} \ 816extern typeout##3 __attribute__((overloadable)) fnc(typein##3 v) { \ 817 typeout##3 r; \ 818 r.x = fnc(v.x); \ 819 r.y = fnc(v.y); \ 820 r.z = fnc(v.z); \ 821 return r; \ 822} \ 823extern typeout##4 __attribute__((overloadable)) fnc(typein##4 v) { \ 824 typeout##4 r; \ 825 r.x = fnc(v.x); \ 826 r.y = fnc(v.y); \ 827 r.z = fnc(v.z); \ 828 r.w = fnc(v.w); \ 829 return r; \ 830} 831 832 833#define UIN_FUNC_IN(fnc) \ 834XN_FUNC_YN(uchar, fnc, char) \ 835XN_FUNC_YN(ushort, fnc, short) \ 836XN_FUNC_YN(uint, fnc, int) 837 838#define IN_FUNC_IN(fnc) \ 839XN_FUNC_YN(uchar, fnc, uchar) \ 840XN_FUNC_YN(char, fnc, char) \ 841XN_FUNC_YN(ushort, fnc, ushort) \ 842XN_FUNC_YN(short, fnc, short) \ 843XN_FUNC_YN(uint, fnc, uint) \ 844XN_FUNC_YN(int, fnc, int) 845 846 847#define XN_FUNC_XN_XN_BODY(type, fnc, body) \ 848extern type __attribute__((overloadable)) \ 849 fnc(type v1, type v2) { \ 850 return body; \ 851} \ 852extern type##2 __attribute__((overloadable)) \ 853 fnc(type##2 v1, type##2 v2) { \ 854 type##2 r; \ 855 r.x = fnc(v1.x, v2.x); \ 856 r.y = fnc(v1.y, v2.y); \ 857 return r; \ 858} \ 859extern type##3 __attribute__((overloadable)) \ 860 fnc(type##3 v1, type##3 v2) { \ 861 type##3 r; \ 862 r.x = fnc(v1.x, v2.x); \ 863 r.y = fnc(v1.y, v2.y); \ 864 r.z = fnc(v1.z, v2.z); \ 865 return r; \ 866} \ 867extern type##4 __attribute__((overloadable)) \ 868 fnc(type##4 v1, type##4 v2) { \ 869 type##4 r; \ 870 r.x = fnc(v1.x, v2.x); \ 871 r.y = fnc(v1.y, v2.y); \ 872 r.z = fnc(v1.z, v2.z); \ 873 r.w = fnc(v1.w, v2.w); \ 874 return r; \ 875} 876 877#define IN_FUNC_IN_IN_BODY(fnc, body) \ 878XN_FUNC_XN_XN_BODY(uchar, fnc, body) \ 879XN_FUNC_XN_XN_BODY(char, fnc, body) \ 880XN_FUNC_XN_XN_BODY(ushort, fnc, body) \ 881XN_FUNC_XN_XN_BODY(short, fnc, body) \ 882XN_FUNC_XN_XN_BODY(uint, fnc, body) \ 883XN_FUNC_XN_XN_BODY(int, fnc, body) \ 884XN_FUNC_XN_XN_BODY(float, fnc, body) 885 886 887/** 888 * abs 889 */ 890extern uint32_t __attribute__((overloadable)) abs(int32_t v) { 891 if (v < 0) 892 return -v; 893 return v; 894} 895extern uint16_t __attribute__((overloadable)) abs(int16_t v) { 896 if (v < 0) 897 return -v; 898 return v; 899} 900extern uint8_t __attribute__((overloadable)) abs(int8_t v) { 901 if (v < 0) 902 return -v; 903 return v; 904} 905 906/** 907 * clz 908 * __builtin_clz only accepts a 32-bit unsigned int, so every input will be 909 * expanded to 32 bits. For our smaller data types, we need to subtract off 910 * these unused top bits (that will be always be composed of zeros). 911 */ 912extern uint32_t __attribute__((overloadable)) clz(uint32_t v) { 913 return __builtin_clz(v); 914} 915extern uint16_t __attribute__((overloadable)) clz(uint16_t v) { 916 return __builtin_clz(v) - 16; 917} 918extern uint8_t __attribute__((overloadable)) clz(uint8_t v) { 919 return __builtin_clz(v) - 24; 920} 921extern int32_t __attribute__((overloadable)) clz(int32_t v) { 922 return __builtin_clz(v); 923} 924extern int16_t __attribute__((overloadable)) clz(int16_t v) { 925 return __builtin_clz(((uint32_t)v) & 0x0000ffff) - 16; 926} 927extern int8_t __attribute__((overloadable)) clz(int8_t v) { 928 return __builtin_clz(((uint32_t)v) & 0x000000ff) - 24; 929} 930 931 932UIN_FUNC_IN(abs) 933IN_FUNC_IN(clz) 934 935 936// 6.11.4 937 938 939extern float __attribute__((overloadable)) degrees(float radians) { 940 return radians * (180.f / M_PI); 941} 942extern float2 __attribute__((overloadable)) degrees(float2 radians) { 943 return radians * (180.f / M_PI); 944} 945extern float3 __attribute__((overloadable)) degrees(float3 radians) { 946 return radians * (180.f / M_PI); 947} 948extern float4 __attribute__((overloadable)) degrees(float4 radians) { 949 return radians * (180.f / M_PI); 950} 951 952extern float __attribute__((overloadable)) mix(float start, float stop, float amount) { 953 return start + (stop - start) * amount; 954} 955extern float2 __attribute__((overloadable)) mix(float2 start, float2 stop, float2 amount) { 956 return start + (stop - start) * amount; 957} 958extern float3 __attribute__((overloadable)) mix(float3 start, float3 stop, float3 amount) { 959 return start + (stop - start) * amount; 960} 961extern float4 __attribute__((overloadable)) mix(float4 start, float4 stop, float4 amount) { 962 return start + (stop - start) * amount; 963} 964extern float2 __attribute__((overloadable)) mix(float2 start, float2 stop, float amount) { 965 return start + (stop - start) * amount; 966} 967extern float3 __attribute__((overloadable)) mix(float3 start, float3 stop, float amount) { 968 return start + (stop - start) * amount; 969} 970extern float4 __attribute__((overloadable)) mix(float4 start, float4 stop, float amount) { 971 return start + (stop - start) * amount; 972} 973 974extern float __attribute__((overloadable)) radians(float degrees) { 975 return degrees * (M_PI / 180.f); 976} 977extern float2 __attribute__((overloadable)) radians(float2 degrees) { 978 return degrees * (M_PI / 180.f); 979} 980extern float3 __attribute__((overloadable)) radians(float3 degrees) { 981 return degrees * (M_PI / 180.f); 982} 983extern float4 __attribute__((overloadable)) radians(float4 degrees) { 984 return degrees * (M_PI / 180.f); 985} 986 987extern float __attribute__((overloadable)) step(float edge, float v) { 988 return (v < edge) ? 0.f : 1.f; 989} 990extern float2 __attribute__((overloadable)) step(float2 edge, float2 v) { 991 float2 r; 992 r.x = (v.x < edge.x) ? 0.f : 1.f; 993 r.y = (v.y < edge.y) ? 0.f : 1.f; 994 return r; 995} 996extern float3 __attribute__((overloadable)) step(float3 edge, float3 v) { 997 float3 r; 998 r.x = (v.x < edge.x) ? 0.f : 1.f; 999 r.y = (v.y < edge.y) ? 0.f : 1.f; 1000 r.z = (v.z < edge.z) ? 0.f : 1.f; 1001 return r; 1002} 1003extern float4 __attribute__((overloadable)) step(float4 edge, float4 v) { 1004 float4 r; 1005 r.x = (v.x < edge.x) ? 0.f : 1.f; 1006 r.y = (v.y < edge.y) ? 0.f : 1.f; 1007 r.z = (v.z < edge.z) ? 0.f : 1.f; 1008 r.w = (v.w < edge.w) ? 0.f : 1.f; 1009 return r; 1010} 1011extern float2 __attribute__((overloadable)) step(float2 edge, float v) { 1012 float2 r; 1013 r.x = (v < edge.x) ? 0.f : 1.f; 1014 r.y = (v < edge.y) ? 0.f : 1.f; 1015 return r; 1016} 1017extern float3 __attribute__((overloadable)) step(float3 edge, float v) { 1018 float3 r; 1019 r.x = (v < edge.x) ? 0.f : 1.f; 1020 r.y = (v < edge.y) ? 0.f : 1.f; 1021 r.z = (v < edge.z) ? 0.f : 1.f; 1022 return r; 1023} 1024extern float4 __attribute__((overloadable)) step(float4 edge, float v) { 1025 float4 r; 1026 r.x = (v < edge.x) ? 0.f : 1.f; 1027 r.y = (v < edge.y) ? 0.f : 1.f; 1028 r.z = (v < edge.z) ? 0.f : 1.f; 1029 r.w = (v < edge.w) ? 0.f : 1.f; 1030 return r; 1031} 1032extern float2 __attribute__((overloadable)) step(float edge, float2 v) { 1033 float2 r; 1034 r.x = (v.x < edge) ? 0.f : 1.f; 1035 r.y = (v.y < edge) ? 0.f : 1.f; 1036 return r; 1037} 1038extern float3 __attribute__((overloadable)) step(float edge, float3 v) { 1039 float3 r; 1040 r.x = (v.x < edge) ? 0.f : 1.f; 1041 r.y = (v.y < edge) ? 0.f : 1.f; 1042 r.z = (v.z < edge) ? 0.f : 1.f; 1043 return r; 1044} 1045extern float4 __attribute__((overloadable)) step(float edge, float4 v) { 1046 float4 r; 1047 r.x = (v.x < edge) ? 0.f : 1.f; 1048 r.y = (v.y < edge) ? 0.f : 1.f; 1049 r.z = (v.z < edge) ? 0.f : 1.f; 1050 r.w = (v.w < edge) ? 0.f : 1.f; 1051 return r; 1052} 1053 1054extern float __attribute__((overloadable)) sign(float v) { 1055 if (v > 0) return 1.f; 1056 if (v < 0) return -1.f; 1057 return v; 1058} 1059FN_FUNC_FN(sign) 1060 1061 1062// 6.11.5 1063extern float3 __attribute__((overloadable)) cross(float3 lhs, float3 rhs) { 1064 float3 r; 1065 r.x = lhs.y * rhs.z - lhs.z * rhs.y; 1066 r.y = lhs.z * rhs.x - lhs.x * rhs.z; 1067 r.z = lhs.x * rhs.y - lhs.y * rhs.x; 1068 return r; 1069} 1070 1071extern float4 __attribute__((overloadable)) cross(float4 lhs, float4 rhs) { 1072 float4 r; 1073 r.x = lhs.y * rhs.z - lhs.z * rhs.y; 1074 r.y = lhs.z * rhs.x - lhs.x * rhs.z; 1075 r.z = lhs.x * rhs.y - lhs.y * rhs.x; 1076 r.w = 0.f; 1077 return r; 1078} 1079 1080#if !defined(ARCH_X86_HAVE_SSSE3) || defined(RS_DEBUG_RUNTIME) || defined(RS_G_RUNTIME) 1081// These functions must be defined here if we are not using the SSE 1082// implementation, which includes when we are built as part of the 1083// debug runtime (libclcore_debug.bc) or compiling with debug info. 1084 1085extern float __attribute__((overloadable)) dot(float lhs, float rhs) { 1086 return lhs * rhs; 1087} 1088extern float __attribute__((overloadable)) dot(float2 lhs, float2 rhs) { 1089 return lhs.x*rhs.x + lhs.y*rhs.y; 1090} 1091extern float __attribute__((overloadable)) dot(float3 lhs, float3 rhs) { 1092 return lhs.x*rhs.x + lhs.y*rhs.y + lhs.z*rhs.z; 1093} 1094extern float __attribute__((overloadable)) dot(float4 lhs, float4 rhs) { 1095 return lhs.x*rhs.x + lhs.y*rhs.y + lhs.z*rhs.z + lhs.w*rhs.w; 1096} 1097 1098extern float __attribute__((overloadable)) length(float v) { 1099 return fabs(v); 1100} 1101extern float __attribute__((overloadable)) length(float2 v) { 1102 return sqrt(v.x*v.x + v.y*v.y); 1103} 1104extern float __attribute__((overloadable)) length(float3 v) { 1105 return sqrt(v.x*v.x + v.y*v.y + v.z*v.z); 1106} 1107extern float __attribute__((overloadable)) length(float4 v) { 1108 return sqrt(v.x*v.x + v.y*v.y + v.z*v.z + v.w*v.w); 1109} 1110 1111#else 1112 1113extern float __attribute__((overloadable)) length(float v); 1114extern float __attribute__((overloadable)) length(float2 v); 1115extern float __attribute__((overloadable)) length(float3 v); 1116extern float __attribute__((overloadable)) length(float4 v); 1117 1118#endif // !defined(ARCH_X86_HAVE_SSSE3) || defined(RS_DEBUG_RUNTIME) || defined(RS_G_RUNTIME) 1119 1120extern float __attribute__((overloadable)) distance(float lhs, float rhs) { 1121 return length(lhs - rhs); 1122} 1123extern float __attribute__((overloadable)) distance(float2 lhs, float2 rhs) { 1124 return length(lhs - rhs); 1125} 1126extern float __attribute__((overloadable)) distance(float3 lhs, float3 rhs) { 1127 return length(lhs - rhs); 1128} 1129extern float __attribute__((overloadable)) distance(float4 lhs, float4 rhs) { 1130 return length(lhs - rhs); 1131} 1132 1133/* For the normalization functions, vectors of length 0 should simply be 1134 * returned (i.e. all the components of that vector are 0). 1135 */ 1136extern float __attribute__((overloadable)) normalize(float v) { 1137 if (v == 0.0f) { 1138 return 0.0f; 1139 } else if (v < 0.0f) { 1140 return -1.0f; 1141 } else { 1142 return 1.0f; 1143 } 1144} 1145extern float2 __attribute__((overloadable)) normalize(float2 v) { 1146 float l = length(v); 1147 return l == 0.0f ? v : v / l; 1148} 1149extern float3 __attribute__((overloadable)) normalize(float3 v) { 1150 float l = length(v); 1151 return l == 0.0f ? v : v / l; 1152} 1153extern float4 __attribute__((overloadable)) normalize(float4 v) { 1154 float l = length(v); 1155 return l == 0.0f ? v : v / l; 1156} 1157 1158extern float __attribute__((overloadable)) half_sqrt(float v) { 1159 return sqrt(v); 1160} 1161FN_FUNC_FN(half_sqrt) 1162 1163extern float __attribute__((overloadable)) fast_length(float v) { 1164 return fabs(v); 1165} 1166extern float __attribute__((overloadable)) fast_length(float2 v) { 1167 return half_sqrt(v.x*v.x + v.y*v.y); 1168} 1169extern float __attribute__((overloadable)) fast_length(float3 v) { 1170 return half_sqrt(v.x*v.x + v.y*v.y + v.z*v.z); 1171} 1172extern float __attribute__((overloadable)) fast_length(float4 v) { 1173 return half_sqrt(v.x*v.x + v.y*v.y + v.z*v.z + v.w*v.w); 1174} 1175 1176extern float __attribute__((overloadable)) fast_distance(float lhs, float rhs) { 1177 return fast_length(lhs - rhs); 1178} 1179extern float __attribute__((overloadable)) fast_distance(float2 lhs, float2 rhs) { 1180 return fast_length(lhs - rhs); 1181} 1182extern float __attribute__((overloadable)) fast_distance(float3 lhs, float3 rhs) { 1183 return fast_length(lhs - rhs); 1184} 1185extern float __attribute__((overloadable)) fast_distance(float4 lhs, float4 rhs) { 1186 return fast_length(lhs - rhs); 1187} 1188 1189extern float __attribute__((overloadable)) half_rsqrt(float); 1190 1191/* For the normalization functions, vectors of length 0 should simply be 1192 * returned (i.e. all the components of that vector are 0). 1193 */ 1194extern float __attribute__((overloadable)) fast_normalize(float v) { 1195 if (v == 0.0f) { 1196 return 0.0f; 1197 } else if (v < 0.0f) { 1198 return -1.0f; 1199 } else { 1200 return 1.0f; 1201 } 1202} 1203// If the length is 0, then rlength should be NaN. 1204extern float2 __attribute__((overloadable)) fast_normalize(float2 v) { 1205 float rlength = half_rsqrt(v.x*v.x + v.y*v.y); 1206 return (rlength == rlength) ? v * rlength : v; 1207} 1208extern float3 __attribute__((overloadable)) fast_normalize(float3 v) { 1209 float rlength = half_rsqrt(v.x*v.x + v.y*v.y + v.z*v.z); 1210 return (rlength == rlength) ? v * rlength : v; 1211} 1212extern float4 __attribute__((overloadable)) fast_normalize(float4 v) { 1213 float rlength = half_rsqrt(v.x*v.x + v.y*v.y + v.z*v.z + v.w*v.w); 1214 return (rlength == rlength) ? v * rlength : v; 1215} 1216 1217extern float __attribute__((overloadable)) half_recip(float v) { 1218 return 1.f / v; 1219} 1220 1221/* 1222extern float __attribute__((overloadable)) approx_atan(float x) { 1223 if (x == 0.f) 1224 return 0.f; 1225 if (x < 0.f) 1226 return -1.f * approx_atan(-1.f * x); 1227 if (x > 1.f) 1228 return M_PI_2 - approx_atan(approx_recip(x)); 1229 return x * approx_recip(1.f + 0.28f * x*x); 1230} 1231FN_FUNC_FN(approx_atan) 1232*/ 1233 1234typedef union 1235{ 1236 float fv; 1237 int32_t iv; 1238} ieee_float_shape_type; 1239 1240/* Get a 32 bit int from a float. */ 1241 1242#define GET_FLOAT_WORD(i,d) \ 1243do { \ 1244 ieee_float_shape_type gf_u; \ 1245 gf_u.fv = (d); \ 1246 (i) = gf_u.iv; \ 1247} while (0) 1248 1249/* Set a float from a 32 bit int. */ 1250 1251#define SET_FLOAT_WORD(d,i) \ 1252do { \ 1253 ieee_float_shape_type sf_u; \ 1254 sf_u.iv = (i); \ 1255 (d) = sf_u.fv; \ 1256} while (0) 1257 1258 1259 1260// Valid -125 to 125 1261extern float __attribute__((overloadable)) native_exp2(float v) { 1262 int32_t iv = (int)v; 1263 int32_t x = iv + (iv >> 31); // ~floor(v) 1264 float r = (v - x); 1265 1266 float fo; 1267 SET_FLOAT_WORD(fo, (x + 127) << 23); 1268 1269 r *= 0.694f; // ~ log(e) / log(2) 1270 float r2 = r*r; 1271 float adj = 1.f + r + (r2 * 0.5f) + (r2*r * 0.166666f) + (r2*r2 * 0.0416666f); 1272 return fo * adj; 1273} 1274 1275extern float2 __attribute__((overloadable)) native_exp2(float2 v) { 1276 int2 iv = convert_int2(v); 1277 int2 x = iv + (iv >> (int2)31);//floor(v); 1278 float2 r = (v - convert_float2(x)); 1279 1280 x += 127; 1281 1282 float2 fo = (float2)(x << (int2)23); 1283 1284 r *= 0.694f; // ~ log(e) / log(2) 1285 float2 r2 = r*r; 1286 float2 adj = 1.f + r + (r2 * 0.5f) + (r2*r * 0.166666f) + (r2*r2 * 0.0416666f); 1287 return fo * adj; 1288} 1289 1290extern float4 __attribute__((overloadable)) native_exp2(float4 v) { 1291 int4 iv = convert_int4(v); 1292 int4 x = iv + (iv >> (int4)31);//floor(v); 1293 float4 r = (v - convert_float4(x)); 1294 1295 x += 127; 1296 1297 float4 fo = (float4)(x << (int4)23); 1298 1299 r *= 0.694f; // ~ log(e) / log(2) 1300 float4 r2 = r*r; 1301 float4 adj = 1.f + r + (r2 * 0.5f) + (r2*r * 0.166666f) + (r2*r2 * 0.0416666f); 1302 return fo * adj; 1303} 1304 1305extern float3 __attribute__((overloadable)) native_exp2(float3 v) { 1306 float4 t = 1.f; 1307 t.xyz = v; 1308 return native_exp2(t).xyz; 1309} 1310 1311 1312extern float __attribute__((overloadable)) native_exp(float v) { 1313 return native_exp2(v * 1.442695041f); 1314} 1315extern float2 __attribute__((overloadable)) native_exp(float2 v) { 1316 return native_exp2(v * 1.442695041f); 1317} 1318extern float3 __attribute__((overloadable)) native_exp(float3 v) { 1319 return native_exp2(v * 1.442695041f); 1320} 1321extern float4 __attribute__((overloadable)) native_exp(float4 v) { 1322 return native_exp2(v * 1.442695041f); 1323} 1324 1325extern float __attribute__((overloadable)) native_exp10(float v) { 1326 return native_exp2(v * 3.321928095f); 1327} 1328extern float2 __attribute__((overloadable)) native_exp10(float2 v) { 1329 return native_exp2(v * 3.321928095f); 1330} 1331extern float3 __attribute__((overloadable)) native_exp10(float3 v) { 1332 return native_exp2(v * 3.321928095f); 1333} 1334extern float4 __attribute__((overloadable)) native_exp10(float4 v) { 1335 return native_exp2(v * 3.321928095f); 1336} 1337 1338extern float __attribute__((overloadable)) native_log2(float v) { 1339 int32_t ibits; 1340 GET_FLOAT_WORD(ibits, v); 1341 1342 int32_t e = (ibits >> 23) & 0xff; 1343 1344 ibits &= 0x7fffff; 1345 ibits |= 127 << 23; 1346 1347 float ir; 1348 SET_FLOAT_WORD(ir, ibits); 1349 ir -= 1.5f; 1350 float ir2 = ir*ir; 1351 float adj2 = (0.405465108f / 0.693147181f) + 1352 ((0.666666667f / 0.693147181f) * ir) - 1353 ((0.222222222f / 0.693147181f) * ir2) + 1354 ((0.098765432f / 0.693147181f) * ir*ir2) - 1355 ((0.049382716f / 0.693147181f) * ir2*ir2) + 1356 ((0.026337449f / 0.693147181f) * ir*ir2*ir2) - 1357 ((0.014631916f / 0.693147181f) * ir2*ir2*ir2); 1358 return (float)(e - 127) + adj2; 1359} 1360extern float2 __attribute__((overloadable)) native_log2(float2 v) { 1361 float2 v2 = {native_log2(v.x), native_log2(v.y)}; 1362 return v2; 1363} 1364extern float3 __attribute__((overloadable)) native_log2(float3 v) { 1365 float3 v2 = {native_log2(v.x), native_log2(v.y), native_log2(v.z)}; 1366 return v2; 1367} 1368extern float4 __attribute__((overloadable)) native_log2(float4 v) { 1369 float4 v2 = {native_log2(v.x), native_log2(v.y), native_log2(v.z), native_log2(v.w)}; 1370 return v2; 1371} 1372 1373extern float __attribute__((overloadable)) native_log(float v) { 1374 return native_log2(v) * (1.f / 1.442695041f); 1375} 1376extern float2 __attribute__((overloadable)) native_log(float2 v) { 1377 return native_log2(v) * (1.f / 1.442695041f); 1378} 1379extern float3 __attribute__((overloadable)) native_log(float3 v) { 1380 return native_log2(v) * (1.f / 1.442695041f); 1381} 1382extern float4 __attribute__((overloadable)) native_log(float4 v) { 1383 return native_log2(v) * (1.f / 1.442695041f); 1384} 1385 1386extern float __attribute__((overloadable)) native_log10(float v) { 1387 return native_log2(v) * (1.f / 3.321928095f); 1388} 1389extern float2 __attribute__((overloadable)) native_log10(float2 v) { 1390 return native_log2(v) * (1.f / 3.321928095f); 1391} 1392extern float3 __attribute__((overloadable)) native_log10(float3 v) { 1393 return native_log2(v) * (1.f / 3.321928095f); 1394} 1395extern float4 __attribute__((overloadable)) native_log10(float4 v) { 1396 return native_log2(v) * (1.f / 3.321928095f); 1397} 1398 1399 1400extern float __attribute__((overloadable)) native_powr(float v, float y) { 1401 float v2 = native_log2(v); 1402 v2 = fmax(v2 * y, -125.f); 1403 return native_exp2(v2); 1404} 1405extern float2 __attribute__((overloadable)) native_powr(float2 v, float2 y) { 1406 float2 v2 = native_log2(v); 1407 v2 = fmax(v2 * y, -125.f); 1408 return native_exp2(v2); 1409} 1410extern float3 __attribute__((overloadable)) native_powr(float3 v, float3 y) { 1411 float3 v2 = native_log2(v); 1412 v2 = fmax(v2 * y, -125.f); 1413 return native_exp2(v2); 1414} 1415extern float4 __attribute__((overloadable)) native_powr(float4 v, float4 y) { 1416 float4 v2 = native_log2(v); 1417 v2 = fmax(v2 * y, -125.f); 1418 return native_exp2(v2); 1419} 1420 1421extern double __attribute__((overloadable)) min(double v1, double v2) { 1422 return v1 < v2 ? v1 : v2; 1423} 1424 1425extern double2 __attribute__((overloadable)) min(double2 v1, double2 v2) { 1426 double2 r; 1427 r.x = v1.x < v2.x ? v1.x : v2.x; 1428 r.y = v1.y < v2.y ? v1.y : v2.y; 1429 return r; 1430} 1431 1432extern double3 __attribute__((overloadable)) min(double3 v1, double3 v2) { 1433 double3 r; 1434 r.x = v1.x < v2.x ? v1.x : v2.x; 1435 r.y = v1.y < v2.y ? v1.y : v2.y; 1436 r.z = v1.z < v2.z ? v1.z : v2.z; 1437 return r; 1438} 1439 1440extern double4 __attribute__((overloadable)) min(double4 v1, double4 v2) { 1441 double4 r; 1442 r.x = v1.x < v2.x ? v1.x : v2.x; 1443 r.y = v1.y < v2.y ? v1.y : v2.y; 1444 r.z = v1.z < v2.z ? v1.z : v2.z; 1445 r.w = v1.w < v2.w ? v1.w : v2.w; 1446 return r; 1447} 1448 1449extern long __attribute__((overloadable)) min(long v1, long v2) { 1450 return v1 < v2 ? v1 : v2; 1451} 1452extern long2 __attribute__((overloadable)) min(long2 v1, long2 v2) { 1453 long2 r; 1454 r.x = v1.x < v2.x ? v1.x : v2.x; 1455 r.y = v1.y < v2.y ? v1.y : v2.y; 1456 return r; 1457} 1458extern long3 __attribute__((overloadable)) min(long3 v1, long3 v2) { 1459 long3 r; 1460 r.x = v1.x < v2.x ? v1.x : v2.x; 1461 r.y = v1.y < v2.y ? v1.y : v2.y; 1462 r.z = v1.z < v2.z ? v1.z : v2.z; 1463 return r; 1464} 1465extern long4 __attribute__((overloadable)) min(long4 v1, long4 v2) { 1466 long4 r; 1467 r.x = v1.x < v2.x ? v1.x : v2.x; 1468 r.y = v1.y < v2.y ? v1.y : v2.y; 1469 r.z = v1.z < v2.z ? v1.z : v2.z; 1470 r.w = v1.w < v2.w ? v1.w : v2.w; 1471 return r; 1472} 1473 1474extern ulong __attribute__((overloadable)) min(ulong v1, ulong v2) { 1475 return v1 < v2 ? v1 : v2; 1476} 1477extern ulong2 __attribute__((overloadable)) min(ulong2 v1, ulong2 v2) { 1478 ulong2 r; 1479 r.x = v1.x < v2.x ? v1.x : v2.x; 1480 r.y = v1.y < v2.y ? v1.y : v2.y; 1481 return r; 1482} 1483extern ulong3 __attribute__((overloadable)) min(ulong3 v1, ulong3 v2) { 1484 ulong3 r; 1485 r.x = v1.x < v2.x ? v1.x : v2.x; 1486 r.y = v1.y < v2.y ? v1.y : v2.y; 1487 r.z = v1.z < v2.z ? v1.z : v2.z; 1488 return r; 1489} 1490extern ulong4 __attribute__((overloadable)) min(ulong4 v1, ulong4 v2) { 1491 ulong4 r; 1492 r.x = v1.x < v2.x ? v1.x : v2.x; 1493 r.y = v1.y < v2.y ? v1.y : v2.y; 1494 r.z = v1.z < v2.z ? v1.z : v2.z; 1495 r.w = v1.w < v2.w ? v1.w : v2.w; 1496 return r; 1497} 1498 1499extern double __attribute__((overloadable)) max(double v1, double v2) { 1500 return v1 > v2 ? v1 : v2; 1501} 1502 1503extern double2 __attribute__((overloadable)) max(double2 v1, double2 v2) { 1504 double2 r; 1505 r.x = v1.x > v2.x ? v1.x : v2.x; 1506 r.y = v1.y > v2.y ? v1.y : v2.y; 1507 return r; 1508} 1509 1510extern double3 __attribute__((overloadable)) max(double3 v1, double3 v2) { 1511 double3 r; 1512 r.x = v1.x > v2.x ? v1.x : v2.x; 1513 r.y = v1.y > v2.y ? v1.y : v2.y; 1514 r.z = v1.z > v2.z ? v1.z : v2.z; 1515 return r; 1516} 1517 1518extern double4 __attribute__((overloadable)) max(double4 v1, double4 v2) { 1519 double4 r; 1520 r.x = v1.x > v2.x ? v1.x : v2.x; 1521 r.y = v1.y > v2.y ? v1.y : v2.y; 1522 r.z = v1.z > v2.z ? v1.z : v2.z; 1523 r.w = v1.w > v2.w ? v1.w : v2.w; 1524 return r; 1525} 1526 1527extern long __attribute__((overloadable)) max(long v1, long v2) { 1528 return v1 > v2 ? v1 : v2; 1529} 1530extern long2 __attribute__((overloadable)) max(long2 v1, long2 v2) { 1531 long2 r; 1532 r.x = v1.x > v2.x ? v1.x : v2.x; 1533 r.y = v1.y > v2.y ? v1.y : v2.y; 1534 return r; 1535} 1536extern long3 __attribute__((overloadable)) max(long3 v1, long3 v2) { 1537 long3 r; 1538 r.x = v1.x > v2.x ? v1.x : v2.x; 1539 r.y = v1.y > v2.y ? v1.y : v2.y; 1540 r.z = v1.z > v2.z ? v1.z : v2.z; 1541 return r; 1542} 1543extern long4 __attribute__((overloadable)) max(long4 v1, long4 v2) { 1544 long4 r; 1545 r.x = v1.x > v2.x ? v1.x : v2.x; 1546 r.y = v1.y > v2.y ? v1.y : v2.y; 1547 r.z = v1.z > v2.z ? v1.z : v2.z; 1548 r.w = v1.w > v2.w ? v1.w : v2.w; 1549 return r; 1550} 1551 1552extern ulong __attribute__((overloadable)) max(ulong v1, ulong v2) { 1553 return v1 > v2 ? v1 : v2; 1554} 1555extern ulong2 __attribute__((overloadable)) max(ulong2 v1, ulong2 v2) { 1556 ulong2 r; 1557 r.x = v1.x > v2.x ? v1.x : v2.x; 1558 r.y = v1.y > v2.y ? v1.y : v2.y; 1559 return r; 1560} 1561extern ulong3 __attribute__((overloadable)) max(ulong3 v1, ulong3 v2) { 1562 ulong3 r; 1563 r.x = v1.x > v2.x ? v1.x : v2.x; 1564 r.y = v1.y > v2.y ? v1.y : v2.y; 1565 r.z = v1.z > v2.z ? v1.z : v2.z; 1566 return r; 1567} 1568extern ulong4 __attribute__((overloadable)) max(ulong4 v1, ulong4 v2) { 1569 ulong4 r; 1570 r.x = v1.x > v2.x ? v1.x : v2.x; 1571 r.y = v1.y > v2.y ? v1.y : v2.y; 1572 r.z = v1.z > v2.z ? v1.z : v2.z; 1573 r.w = v1.w > v2.w ? v1.w : v2.w; 1574 return r; 1575} 1576 1577#define THUNK_NATIVE_F(fn) \ 1578 float __attribute__((overloadable)) native_##fn(float v) { return fn(v);} \ 1579 float2 __attribute__((overloadable)) native_##fn(float2 v) { return fn(v);} \ 1580 float3 __attribute__((overloadable)) native_##fn(float3 v) { return fn(v);} \ 1581 float4 __attribute__((overloadable)) native_##fn(float4 v) { return fn(v);} 1582 1583#define THUNK_NATIVE_F_F(fn) \ 1584 float __attribute__((overloadable)) native_##fn(float v1, float v2) { return fn(v1, v2);} \ 1585 float2 __attribute__((overloadable)) native_##fn(float2 v1, float2 v2) { return fn(v1, v2);} \ 1586 float3 __attribute__((overloadable)) native_##fn(float3 v1, float3 v2) { return fn(v1, v2);} \ 1587 float4 __attribute__((overloadable)) native_##fn(float4 v1, float4 v2) { return fn(v1, v2);} 1588 1589#define THUNK_NATIVE_F_FP(fn) \ 1590 float __attribute__((overloadable)) native_##fn(float v1, float *v2) { return fn(v1, v2);} \ 1591 float2 __attribute__((overloadable)) native_##fn(float2 v1, float2 *v2) { return fn(v1, v2);} \ 1592 float3 __attribute__((overloadable)) native_##fn(float3 v1, float3 *v2) { return fn(v1, v2);} \ 1593 float4 __attribute__((overloadable)) native_##fn(float4 v1, float4 *v2) { return fn(v1, v2);} 1594 1595#define THUNK_NATIVE_F_I(fn) \ 1596 float __attribute__((overloadable)) native_##fn(float v1, int v2) { return fn(v1, v2);} \ 1597 float2 __attribute__((overloadable)) native_##fn(float2 v1, int2 v2) { return fn(v1, v2);} \ 1598 float3 __attribute__((overloadable)) native_##fn(float3 v1, int3 v2) { return fn(v1, v2);} \ 1599 float4 __attribute__((overloadable)) native_##fn(float4 v1, int4 v2) { return fn(v1, v2);} 1600 1601THUNK_NATIVE_F(acos) 1602THUNK_NATIVE_F(acosh) 1603THUNK_NATIVE_F(acospi) 1604THUNK_NATIVE_F(asin) 1605THUNK_NATIVE_F(asinh) 1606THUNK_NATIVE_F(asinpi) 1607THUNK_NATIVE_F(atan) 1608THUNK_NATIVE_F_F(atan2) 1609THUNK_NATIVE_F(atanh) 1610THUNK_NATIVE_F(atanpi) 1611THUNK_NATIVE_F_F(atan2pi) 1612THUNK_NATIVE_F(cbrt) 1613THUNK_NATIVE_F(cos) 1614THUNK_NATIVE_F(cosh) 1615THUNK_NATIVE_F(cospi) 1616THUNK_NATIVE_F(expm1) 1617THUNK_NATIVE_F_F(hypot) 1618THUNK_NATIVE_F(log1p) 1619THUNK_NATIVE_F_I(rootn) 1620THUNK_NATIVE_F(rsqrt) 1621THUNK_NATIVE_F(sqrt) 1622THUNK_NATIVE_F(sin) 1623THUNK_NATIVE_F_FP(sincos) 1624THUNK_NATIVE_F(sinh) 1625THUNK_NATIVE_F(sinpi) 1626THUNK_NATIVE_F(tan) 1627THUNK_NATIVE_F(tanh) 1628THUNK_NATIVE_F(tanpi) 1629 1630#undef THUNK_NATIVE_F 1631#undef THUNK_NATIVE_F_F 1632#undef THUNK_NATIVE_F_I 1633#undef THUNK_NATIVE_F_FP 1634 1635float __attribute__((overloadable)) native_normalize(float v) { return fast_normalize(v);} 1636float2 __attribute__((overloadable)) native_normalize(float2 v) { return fast_normalize(v);} 1637float3 __attribute__((overloadable)) native_normalize(float3 v) { return fast_normalize(v);} 1638float4 __attribute__((overloadable)) native_normalize(float4 v) { return fast_normalize(v);} 1639 1640float __attribute__((overloadable)) native_distance(float v1, float v2) { return fast_distance(v1, v2);} 1641float __attribute__((overloadable)) native_distance(float2 v1, float2 v2) { return fast_distance(v1, v2);} 1642float __attribute__((overloadable)) native_distance(float3 v1, float3 v2) { return fast_distance(v1, v2);} 1643float __attribute__((overloadable)) native_distance(float4 v1, float4 v2) { return fast_distance(v1, v2);} 1644 1645float __attribute__((overloadable)) native_length(float v) { return fast_length(v);} 1646float __attribute__((overloadable)) native_length(float2 v) { return fast_length(v);} 1647float __attribute__((overloadable)) native_length(float3 v) { return fast_length(v);} 1648float __attribute__((overloadable)) native_length(float4 v) { return fast_length(v);} 1649 1650float __attribute__((overloadable)) native_divide(float v1, float v2) { return v1 / v2;} 1651float2 __attribute__((overloadable)) native_divide(float2 v1, float2 v2) { return v1 / v2;} 1652float3 __attribute__((overloadable)) native_divide(float3 v1, float3 v2) { return v1 / v2;} 1653float4 __attribute__((overloadable)) native_divide(float4 v1, float4 v2) { return v1 / v2;} 1654 1655float __attribute__((overloadable)) native_recip(float v) { return 1.f / v;} 1656float2 __attribute__((overloadable)) native_recip(float2 v) { return ((float2)1.f) / v;} 1657float3 __attribute__((overloadable)) native_recip(float3 v) { return ((float3)1.f) / v;} 1658float4 __attribute__((overloadable)) native_recip(float4 v) { return ((float4)1.f) / v;} 1659 1660 1661 1662 1663 1664#undef FN_FUNC_FN 1665#undef IN_FUNC_FN 1666#undef FN_FUNC_FN_FN 1667#undef FN_FUNC_FN_F 1668#undef FN_FUNC_FN_IN 1669#undef FN_FUNC_FN_I 1670#undef FN_FUNC_FN_PFN 1671#undef FN_FUNC_FN_PIN 1672#undef FN_FUNC_FN_FN_FN 1673#undef FN_FUNC_FN_FN_PIN 1674#undef XN_FUNC_YN 1675#undef UIN_FUNC_IN 1676#undef IN_FUNC_IN 1677#undef XN_FUNC_XN_XN_BODY 1678#undef IN_FUNC_IN_IN_BODY 1679 1680static const unsigned short kHalfPositiveInfinity = 0x7c00; 1681 1682/* Define f16 functions of the form 1683 * HN output = fn(HN input) 1684 * where HN is scalar or vector half type 1685 */ 1686#define HN_FUNC_HN(fn) \ 1687extern half __attribute__((overloadable)) fn(half h) { \ 1688 return (half) fn((float) h); \ 1689} \ 1690extern half2 __attribute__((overloadable)) fn(half2 v) { \ 1691 return convert_half2(fn(convert_float2(v))); \ 1692} \ 1693extern half3 __attribute__((overloadable)) fn(half3 v) { \ 1694 return convert_half3(fn(convert_float3(v))); \ 1695} \ 1696extern half4 __attribute__((overloadable)) fn(half4 v) { \ 1697 return convert_half4(fn(convert_float4(v))); \ 1698} 1699 1700/* Define f16 functions of the form 1701 * HN output = fn(HN input1, HN input2) 1702 * where HN is scalar or vector half type 1703 */ 1704#define HN_FUNC_HN_HN(fn) \ 1705extern half __attribute__((overloadable)) fn(half h1, half h2) { \ 1706 return (half) fn((float) h1, (float) h2); \ 1707} \ 1708extern half2 __attribute__((overloadable)) fn(half2 v1, half2 v2) { \ 1709 return convert_half2(fn(convert_float2(v1), \ 1710 convert_float2(v2))); \ 1711} \ 1712extern half3 __attribute__((overloadable)) fn(half3 v1, half3 v2) { \ 1713 return convert_half3(fn(convert_float3(v1), \ 1714 convert_float3(v2))); \ 1715} \ 1716extern half4 __attribute__((overloadable)) fn(half4 v1, half4 v2) { \ 1717 return convert_half4(fn(convert_float4(v1), \ 1718 convert_float4(v2))); \ 1719} 1720 1721/* Define f16 functions of the form 1722 * HN output = fn(HN input1, half input2) 1723 * where HN is scalar or vector half type 1724 */ 1725#define HN_FUNC_HN_H(fn) \ 1726extern half2 __attribute__((overloadable)) fn(half2 v1, half v2) { \ 1727 return convert_half2(fn(convert_float2(v1), (float) v2)); \ 1728} \ 1729extern half3 __attribute__((overloadable)) fn(half3 v1, half v2) { \ 1730 return convert_half3(fn(convert_float3(v1), (float) v2)); \ 1731} \ 1732extern half4 __attribute__((overloadable)) fn(half4 v1, half v2) { \ 1733 return convert_half4(fn(convert_float4(v1), (float) v2)); \ 1734} 1735 1736/* Define f16 functions of the form 1737 * HN output = fn(HN input1, HN input2, HN input3) 1738 * where HN is scalar or vector half type 1739 */ 1740#define HN_FUNC_HN_HN_HN(fn) \ 1741extern half __attribute__((overloadable)) fn(half h1, half h2, half h3) { \ 1742 return (half) fn((float) h1, (float) h2, (float) h3); \ 1743} \ 1744extern half2 __attribute__((overloadable)) fn(half2 v1, half2 v2, half2 v3) { \ 1745 return convert_half2(fn(convert_float2(v1), \ 1746 convert_float2(v2), \ 1747 convert_float2(v3))); \ 1748} \ 1749extern half3 __attribute__((overloadable)) fn(half3 v1, half3 v2, half3 v3) { \ 1750 return convert_half3(fn(convert_float3(v1), \ 1751 convert_float3(v2), \ 1752 convert_float3(v3))); \ 1753} \ 1754extern half4 __attribute__((overloadable)) fn(half4 v1, half4 v2, half4 v3) { \ 1755 return convert_half4(fn(convert_float4(v1), \ 1756 convert_float4(v2), \ 1757 convert_float4(v3))); \ 1758} 1759 1760/* Define f16 functions of the form 1761 * HN output = fn(HN input1, IN input2) 1762 * where HN is scalar or vector half type and IN the equivalent integer type 1763 * of same vector length. 1764 */ 1765#define HN_FUNC_HN_IN(fn) \ 1766extern half __attribute__((overloadable)) fn(half h1, int v) { \ 1767 return (half) fn((float) h1, v); \ 1768} \ 1769extern half2 __attribute__((overloadable)) fn(half2 v1, int2 v2) { \ 1770 return convert_half2(fn(convert_float2(v1), v2)); \ 1771} \ 1772extern half3 __attribute__((overloadable)) fn(half3 v1, int3 v2) { \ 1773 return convert_half3(fn(convert_float3(v1), v2)); \ 1774} \ 1775extern half4 __attribute__((overloadable)) fn(half4 v1, int4 v2) { \ 1776 return convert_half4(fn(convert_float4(v1), v2)); \ 1777} 1778 1779/* Define f16 functions of the form 1780 * half output = fn(HN input1) 1781 * where HN is a scalar or vector half type. 1782 */ 1783#define H_FUNC_HN(fn) \ 1784extern half __attribute__((overloadable)) fn(half h) { \ 1785 return (half) fn((float) h); \ 1786} \ 1787extern half __attribute__((overloadable)) fn(half2 v) { \ 1788 return fn(convert_float2(v)); \ 1789} \ 1790extern half __attribute__((overloadable)) fn(half3 v) { \ 1791 return fn(convert_float3(v)); \ 1792} \ 1793extern half __attribute__((overloadable)) fn(half4 v) { \ 1794 return fn(convert_float4(v)); \ 1795} 1796 1797/* Define f16 functions of the form 1798 * half output = fn(HN input1, HN input2) 1799 * where HN is a scalar or vector half type. 1800 */ 1801#define H_FUNC_HN_HN(fn) \ 1802extern half __attribute__((overloadable)) fn(half h1, half h2) { \ 1803 return (half) fn((float) h1, (float) h2); \ 1804} \ 1805extern half __attribute__((overloadable)) fn(half2 v1, half2 v2) { \ 1806 return fn(convert_float2(v1), convert_float2(v2)); \ 1807} \ 1808extern half __attribute__((overloadable)) fn(half3 v1, half3 v2) { \ 1809 return fn(convert_float3(v1), convert_float3(v2)); \ 1810} \ 1811extern half __attribute__((overloadable)) fn(half4 v1, half4 v2) { \ 1812 return fn(convert_float4(v1), convert_float4(v2)); \ 1813} 1814 1815#define SCALARIZE_HN_FUNC_HN_PHN(fnc) \ 1816extern half2 __attribute__((overloadable)) fnc(half2 v1, half2 *v2) { \ 1817 half2 ret; \ 1818 half t[2]; \ 1819 ret.x = fnc(v1.x, &t[0]); \ 1820 ret.y = fnc(v1.y, &t[1]); \ 1821 v2->x = t[0]; \ 1822 v2->y = t[1]; \ 1823 return ret; \ 1824} \ 1825extern half3 __attribute__((overloadable)) fnc(half3 v1, half3 *v2) { \ 1826 half3 ret; \ 1827 half t[3]; \ 1828 ret.x = fnc(v1.x, &t[0]); \ 1829 ret.y = fnc(v1.y, &t[1]); \ 1830 ret.z = fnc(v1.z, &t[2]); \ 1831 v2->x = t[0]; \ 1832 v2->y = t[1]; \ 1833 v2->z = t[2]; \ 1834 return ret; \ 1835} \ 1836extern half4 __attribute__((overloadable)) fnc(half4 v1, half4 *v2) { \ 1837 half4 ret; \ 1838 half t[4]; \ 1839 ret.x = fnc(v1.x, &t[0]); \ 1840 ret.y = fnc(v1.y, &t[1]); \ 1841 ret.z = fnc(v1.z, &t[2]); \ 1842 ret.w = fnc(v1.w, &t[3]); \ 1843 v2->x = t[0]; \ 1844 v2->y = t[1]; \ 1845 v2->z = t[2]; \ 1846 v2->w = t[3]; \ 1847 return ret; \ 1848} 1849 1850/* Define f16 functions of the form 1851 * HN output = fn(HN input1, HN input2) 1852 * where HN is a vector half type. The functions are defined to call the 1853 * scalar function of the same name. 1854 */ 1855#define SCALARIZE_HN_FUNC_HN_HN(fn) \ 1856extern half2 __attribute__((overloadable)) fn(half2 v1, half2 v2) { \ 1857 half2 ret; \ 1858 ret.x = fn(v1.x, v2.x); \ 1859 ret.y = fn(v1.y, v2.y); \ 1860 return ret; \ 1861} \ 1862extern half3 __attribute__((overloadable)) fn(half3 v1, half3 v2) { \ 1863 half3 ret; \ 1864 ret.x = fn(v1.x, v2.x); \ 1865 ret.y = fn(v1.y, v2.y); \ 1866 ret.z = fn(v1.z, v2.z); \ 1867 return ret; \ 1868} \ 1869extern half4 __attribute__((overloadable)) fn(half4 v1, half4 v2) { \ 1870 half4 ret; \ 1871 ret.x = fn(v1.x, v2.x); \ 1872 ret.y = fn(v1.y, v2.y); \ 1873 ret.z = fn(v1.z, v2.z); \ 1874 ret.w = fn(v1.w, v2.w); \ 1875 return ret; \ 1876} \ 1877 1878HN_FUNC_HN(acos); 1879HN_FUNC_HN(acosh); 1880HN_FUNC_HN(acospi); 1881HN_FUNC_HN(asin); 1882HN_FUNC_HN(asinh); 1883HN_FUNC_HN(asinpi); 1884HN_FUNC_HN(atan); 1885HN_FUNC_HN(atanh); 1886HN_FUNC_HN(atanpi); 1887HN_FUNC_HN_HN(atan2); 1888HN_FUNC_HN_HN(atan2pi); 1889 1890HN_FUNC_HN(cbrt); 1891HN_FUNC_HN(ceil); 1892 1893extern half __attribute__((overloadable)) copysign(half x, half y); 1894SCALARIZE_HN_FUNC_HN_HN(copysign); 1895 1896HN_FUNC_HN(cos); 1897HN_FUNC_HN(cosh); 1898HN_FUNC_HN(cospi); 1899 1900extern half3 __attribute__((overloadable)) cross(half3 lhs, half3 rhs) { 1901 half3 r; 1902 r.x = lhs.y * rhs.z - lhs.z * rhs.y; 1903 r.y = lhs.z * rhs.x - lhs.x * rhs.z; 1904 r.z = lhs.x * rhs.y - lhs.y * rhs.x; 1905 return r; 1906} 1907 1908extern half4 __attribute__((overloadable)) cross(half4 lhs, half4 rhs) { 1909 half4 r; 1910 r.x = lhs.y * rhs.z - lhs.z * rhs.y; 1911 r.y = lhs.z * rhs.x - lhs.x * rhs.z; 1912 r.z = lhs.x * rhs.y - lhs.y * rhs.x; 1913 r.w = 0.f; 1914 return r; 1915} 1916 1917HN_FUNC_HN(degrees); 1918H_FUNC_HN_HN(distance); 1919H_FUNC_HN_HN(dot); 1920 1921HN_FUNC_HN(erf); 1922HN_FUNC_HN(erfc); 1923HN_FUNC_HN(exp); 1924HN_FUNC_HN(exp10); 1925HN_FUNC_HN(exp2); 1926HN_FUNC_HN(expm1); 1927 1928HN_FUNC_HN(fabs); 1929HN_FUNC_HN_HN(fdim); 1930HN_FUNC_HN(floor); 1931HN_FUNC_HN_HN_HN(fma); 1932HN_FUNC_HN_HN(fmax); 1933HN_FUNC_HN_H(fmax); 1934HN_FUNC_HN_HN(fmin); 1935HN_FUNC_HN_H(fmin); 1936HN_FUNC_HN_HN(fmod); 1937 1938extern half __attribute__((overloadable)) fract(half v, half *iptr) { 1939 // maxLessThanOne = 0.99951171875, the largest value < 1.0 1940 half maxLessThanOne; 1941 SET_HALF_WORD(maxLessThanOne, 0x3bff); 1942 1943 int i = (int) floor(v); 1944 if (iptr) { 1945 *iptr = i; 1946 } 1947 // return v - floor(v), if strictly less than one 1948 return fmin(v - i, maxLessThanOne); 1949} 1950 1951SCALARIZE_HN_FUNC_HN_PHN(fract); 1952 1953extern half __attribute__((const, overloadable)) fract(half v) { 1954 half unused; 1955 return fract(v, &unused); 1956} 1957 1958extern half2 __attribute__((const, overloadable)) fract(half2 v) { 1959 half2 unused; 1960 return fract(v, &unused); 1961} 1962 1963extern half3 __attribute__((const, overloadable)) fract(half3 v) { 1964 half3 unused; 1965 return fract(v, &unused); 1966} 1967 1968extern half4 __attribute__((const, overloadable)) fract(half4 v) { 1969 half4 unused; 1970 return fract(v, &unused); 1971} 1972 1973extern half __attribute__((overloadable)) frexp(half x, int *eptr); 1974 1975extern half2 __attribute__((overloadable)) frexp(half2 v1, int2 *eptr) { 1976 half2 ret; 1977 int e[2]; 1978 ret.x = frexp(v1.x, &e[0]); 1979 ret.y = frexp(v1.y, &e[1]); 1980 eptr->x = e[0]; 1981 eptr->y = e[1]; 1982 return ret; 1983} 1984 1985extern half3 __attribute__((overloadable)) frexp(half3 v1, int3 *eptr) { 1986 half3 ret; 1987 int e[3]; 1988 ret.x = frexp(v1.x, &e[0]); 1989 ret.y = frexp(v1.y, &e[1]); 1990 ret.z = frexp(v1.z, &e[2]); 1991 eptr->x = e[0]; 1992 eptr->y = e[1]; 1993 eptr->z = e[2]; 1994 return ret; 1995} 1996 1997extern half4 __attribute__((overloadable)) frexp(half4 v1, int4 *eptr) { 1998 half4 ret; 1999 int e[4]; 2000 ret.x = frexp(v1.x, &e[0]); 2001 ret.y = frexp(v1.y, &e[1]); 2002 ret.z = frexp(v1.z, &e[2]); 2003 ret.w = frexp(v1.w, &e[3]); 2004 eptr->x = e[0]; 2005 eptr->y = e[1]; 2006 eptr->z = e[2]; 2007 eptr->w = e[3]; 2008 return ret; 2009} 2010 2011HN_FUNC_HN_HN(hypot); 2012 2013extern int __attribute__((overloadable)) ilogb(half x); 2014 2015extern int2 __attribute__((overloadable)) ilogb(half2 v) { 2016 int2 ret; 2017 ret.x = ilogb(v.x); 2018 ret.y = ilogb(v.y); 2019 return ret; 2020} 2021extern int3 __attribute__((overloadable)) ilogb(half3 v) { 2022 int3 ret; 2023 ret.x = ilogb(v.x); 2024 ret.y = ilogb(v.y); 2025 ret.z = ilogb(v.z); 2026 return ret; 2027} 2028extern int4 __attribute__((overloadable)) ilogb(half4 v) { 2029 int4 ret; 2030 ret.x = ilogb(v.x); 2031 ret.y = ilogb(v.y); 2032 ret.z = ilogb(v.z); 2033 ret.w = ilogb(v.w); 2034 return ret; 2035} 2036 2037HN_FUNC_HN_IN(ldexp); 2038extern half2 __attribute__((overloadable)) ldexp(half2 v, int exponent) { 2039 return convert_half2(ldexp(convert_float2(v), exponent)); 2040} 2041extern half3 __attribute__((overloadable)) ldexp(half3 v, int exponent) { 2042 return convert_half3(ldexp(convert_float3(v), exponent)); 2043} 2044extern half4 __attribute__((overloadable)) ldexp(half4 v, int exponent) { 2045 return convert_half4(ldexp(convert_float4(v), exponent)); 2046} 2047 2048H_FUNC_HN(length); 2049HN_FUNC_HN(lgamma); 2050 2051extern half __attribute__((overloadable)) lgamma(half h, int *signp) { 2052 return (half) lgamma((float) h, signp); 2053} 2054extern half2 __attribute__((overloadable)) lgamma(half2 v, int2 *signp) { 2055 return convert_half2(lgamma(convert_float2(v), signp)); 2056} 2057extern half3 __attribute__((overloadable)) lgamma(half3 v, int3 *signp) { 2058 return convert_half3(lgamma(convert_float3(v), signp)); 2059} 2060extern half4 __attribute__((overloadable)) lgamma(half4 v, int4 *signp) { 2061 return convert_half4(lgamma(convert_float4(v), signp)); 2062} 2063 2064HN_FUNC_HN(log); 2065HN_FUNC_HN(log10); 2066HN_FUNC_HN(log1p); 2067HN_FUNC_HN(log2); 2068HN_FUNC_HN(logb); 2069 2070HN_FUNC_HN_HN_HN(mad); 2071HN_FUNC_HN_HN(max); 2072HN_FUNC_HN_H(max); // TODO can this be arch-specific similar to _Z3maxDv2_ff? 2073HN_FUNC_HN_HN(min); 2074HN_FUNC_HN_H(min); // TODO can this be arch-specific similar to _Z3minDv2_ff? 2075 2076extern half __attribute__((overloadable)) mix(half start, half stop, half amount) { 2077 return start + (stop - start) * amount; 2078} 2079extern half2 __attribute__((overloadable)) mix(half2 start, half2 stop, half2 amount) { 2080 return start + (stop - start) * amount; 2081} 2082extern half3 __attribute__((overloadable)) mix(half3 start, half3 stop, half3 amount) { 2083 return start + (stop - start) * amount; 2084} 2085extern half4 __attribute__((overloadable)) mix(half4 start, half4 stop, half4 amount) { 2086 return start + (stop - start) * amount; 2087} 2088extern half2 __attribute__((overloadable)) mix(half2 start, half2 stop, half amount) { 2089 return start + (stop - start) * amount; 2090} 2091extern half3 __attribute__((overloadable)) mix(half3 start, half3 stop, half amount) { 2092 return start + (stop - start) * amount; 2093} 2094extern half4 __attribute__((overloadable)) mix(half4 start, half4 stop, half amount) { 2095 return start + (stop - start) * amount; 2096} 2097 2098extern half __attribute__((overloadable)) modf(half x, half *iptr); 2099SCALARIZE_HN_FUNC_HN_PHN(modf); 2100 2101half __attribute__((overloadable)) nan_half() { 2102 unsigned short nan_short = kHalfPositiveInfinity | 0x0200; 2103 half nan; 2104 SET_HALF_WORD(nan, nan_short); 2105 return nan; 2106} 2107 2108HN_FUNC_HN(normalize); 2109 2110extern half __attribute__((overloadable)) nextafter(half x, half y); 2111SCALARIZE_HN_FUNC_HN_HN(nextafter); 2112 2113HN_FUNC_HN_HN(pow); 2114HN_FUNC_HN_IN(pown); 2115HN_FUNC_HN_HN(powr); 2116HN_FUNC_HN(radians); 2117HN_FUNC_HN_HN(remainder); 2118 2119extern half __attribute__((overloadable)) remquo(half n, half d, int *quo) { 2120 return (float) remquo((float) n, (float) d, quo); 2121} 2122extern half2 __attribute__((overloadable)) remquo(half2 n, half2 d, int2 *quo) { 2123 return convert_half2(remquo(convert_float2(d), convert_float2(n), quo)); 2124} 2125extern half3 __attribute__((overloadable)) remquo(half3 n, half3 d, int3 *quo) { 2126 return convert_half3(remquo(convert_float3(d), convert_float3(n), quo)); 2127} 2128extern half4 __attribute__((overloadable)) remquo(half4 n, half4 d, int4 *quo) { 2129 return convert_half4(remquo(convert_float4(d), convert_float4(n), quo)); 2130} 2131 2132HN_FUNC_HN(rint); 2133HN_FUNC_HN_IN(rootn); 2134HN_FUNC_HN(round); 2135HN_FUNC_HN(rsqrt); 2136 2137extern half __attribute__((overloadable)) sign(half h) { 2138 if (h > 0) return (half) 1.f; 2139 if (h < 0) return (half) -1.f; 2140 return h; 2141} 2142extern half2 __attribute__((overloadable)) sign(half2 v) { 2143 half2 ret; 2144 ret.x = sign(v.x); 2145 ret.y = sign(v.y); 2146 return ret; 2147} 2148extern half3 __attribute__((overloadable)) sign(half3 v) { 2149 half3 ret; 2150 ret.x = sign(v.x); 2151 ret.y = sign(v.y); 2152 ret.z = sign(v.z); 2153 return ret; 2154} 2155extern half4 __attribute__((overloadable)) sign(half4 v) { 2156 half4 ret; 2157 ret.x = sign(v.x); 2158 ret.y = sign(v.y); 2159 ret.z = sign(v.z); 2160 ret.w = sign(v.w); 2161 return ret; 2162} 2163 2164HN_FUNC_HN(sin); 2165 2166extern half __attribute__((overloadable)) sincos(half v, half *cosptr) { 2167 *cosptr = cos(v); 2168 return sin(v); 2169} 2170// TODO verify if LLVM eliminates the duplicate convert_float2 2171extern half2 __attribute__((overloadable)) sincos(half2 v, half2 *cosptr) { 2172 *cosptr = cos(v); 2173 return sin(v); 2174} 2175extern half3 __attribute__((overloadable)) sincos(half3 v, half3 *cosptr) { 2176 *cosptr = cos(v); 2177 return sin(v); 2178} 2179extern half4 __attribute__((overloadable)) sincos(half4 v, half4 *cosptr) { 2180 *cosptr = cos(v); 2181 return sin(v); 2182} 2183 2184HN_FUNC_HN(sinh); 2185HN_FUNC_HN(sinpi); 2186HN_FUNC_HN(sqrt); 2187 2188extern half __attribute__((overloadable)) step(half edge, half v) { 2189 return (v < edge) ? 0.f : 1.f; 2190} 2191extern half2 __attribute__((overloadable)) step(half2 edge, half2 v) { 2192 half2 r; 2193 r.x = (v.x < edge.x) ? 0.f : 1.f; 2194 r.y = (v.y < edge.y) ? 0.f : 1.f; 2195 return r; 2196} 2197extern half3 __attribute__((overloadable)) step(half3 edge, half3 v) { 2198 half3 r; 2199 r.x = (v.x < edge.x) ? 0.f : 1.f; 2200 r.y = (v.y < edge.y) ? 0.f : 1.f; 2201 r.z = (v.z < edge.z) ? 0.f : 1.f; 2202 return r; 2203} 2204extern half4 __attribute__((overloadable)) step(half4 edge, half4 v) { 2205 half4 r; 2206 r.x = (v.x < edge.x) ? 0.f : 1.f; 2207 r.y = (v.y < edge.y) ? 0.f : 1.f; 2208 r.z = (v.z < edge.z) ? 0.f : 1.f; 2209 r.w = (v.w < edge.w) ? 0.f : 1.f; 2210 return r; 2211} 2212extern half2 __attribute__((overloadable)) step(half2 edge, half v) { 2213 half2 r; 2214 r.x = (v < edge.x) ? 0.f : 1.f; 2215 r.y = (v < edge.y) ? 0.f : 1.f; 2216 return r; 2217} 2218extern half3 __attribute__((overloadable)) step(half3 edge, half v) { 2219 half3 r; 2220 r.x = (v < edge.x) ? 0.f : 1.f; 2221 r.y = (v < edge.y) ? 0.f : 1.f; 2222 r.z = (v < edge.z) ? 0.f : 1.f; 2223 return r; 2224} 2225extern half4 __attribute__((overloadable)) step(half4 edge, half v) { 2226 half4 r; 2227 r.x = (v < edge.x) ? 0.f : 1.f; 2228 r.y = (v < edge.y) ? 0.f : 1.f; 2229 r.z = (v < edge.z) ? 0.f : 1.f; 2230 r.w = (v < edge.w) ? 0.f : 1.f; 2231 return r; 2232} 2233extern half2 __attribute__((overloadable)) step(half edge, half2 v) { 2234 half2 r; 2235 r.x = (v.x < edge) ? 0.f : 1.f; 2236 r.y = (v.y < edge) ? 0.f : 1.f; 2237 return r; 2238} 2239extern half3 __attribute__((overloadable)) step(half edge, half3 v) { 2240 half3 r; 2241 r.x = (v.x < edge) ? 0.f : 1.f; 2242 r.y = (v.y < edge) ? 0.f : 1.f; 2243 r.z = (v.z < edge) ? 0.f : 1.f; 2244 return r; 2245} 2246extern half4 __attribute__((overloadable)) step(half edge, half4 v) { 2247 half4 r; 2248 r.x = (v.x < edge) ? 0.f : 1.f; 2249 r.y = (v.y < edge) ? 0.f : 1.f; 2250 r.z = (v.z < edge) ? 0.f : 1.f; 2251 r.w = (v.w < edge) ? 0.f : 1.f; 2252 return r; 2253} 2254 2255HN_FUNC_HN(tan); 2256HN_FUNC_HN(tanh); 2257HN_FUNC_HN(tanpi); 2258HN_FUNC_HN(tgamma); 2259HN_FUNC_HN(trunc); // TODO: rethink: needs half-specific implementation? 2260 2261HN_FUNC_HN(native_acos); 2262HN_FUNC_HN(native_acosh); 2263HN_FUNC_HN(native_acospi); 2264HN_FUNC_HN(native_asin); 2265HN_FUNC_HN(native_asinh); 2266HN_FUNC_HN(native_asinpi); 2267HN_FUNC_HN(native_atan); 2268HN_FUNC_HN(native_atanh); 2269HN_FUNC_HN(native_atanpi); 2270HN_FUNC_HN_HN(native_atan2); 2271HN_FUNC_HN_HN(native_atan2pi); 2272 2273HN_FUNC_HN(native_cbrt); 2274HN_FUNC_HN(native_cos); 2275HN_FUNC_HN(native_cosh); 2276HN_FUNC_HN(native_cospi); 2277 2278H_FUNC_HN_HN(native_distance); 2279HN_FUNC_HN_HN(native_divide); 2280 2281HN_FUNC_HN(native_exp); 2282HN_FUNC_HN(native_exp10); 2283HN_FUNC_HN(native_exp2); 2284HN_FUNC_HN(native_expm1); 2285 2286HN_FUNC_HN_HN(native_hypot); 2287H_FUNC_HN(native_length); 2288 2289HN_FUNC_HN(native_log); 2290HN_FUNC_HN(native_log10); 2291HN_FUNC_HN(native_log1p); 2292HN_FUNC_HN(native_log2); 2293 2294HN_FUNC_HN(native_normalize); 2295 2296HN_FUNC_HN_HN(native_powr); // TODO are parameter limits different for half? 2297 2298HN_FUNC_HN(native_recip); 2299HN_FUNC_HN_IN(native_rootn); 2300HN_FUNC_HN(native_rsqrt); 2301 2302HN_FUNC_HN(native_sin); 2303 2304extern half __attribute__((overloadable)) native_sincos(half v, half *cosptr) { 2305 return sincos(v, cosptr); 2306} 2307extern half2 __attribute__((overloadable)) native_sincos(half2 v, half2 *cosptr) { 2308 return sincos(v, cosptr); 2309} 2310extern half3 __attribute__((overloadable)) native_sincos(half3 v, half3 *cosptr) { 2311 return sincos(v, cosptr); 2312} 2313extern half4 __attribute__((overloadable)) native_sincos(half4 v, half4 *cosptr) { 2314 return sincos(v, cosptr); 2315} 2316 2317HN_FUNC_HN(native_sinh); 2318HN_FUNC_HN(native_sinpi); 2319HN_FUNC_HN(native_sqrt); 2320 2321HN_FUNC_HN(native_tan); 2322HN_FUNC_HN(native_tanh); 2323HN_FUNC_HN(native_tanpi); 2324 2325#undef HN_FUNC_HN 2326#undef HN_FUNC_HN_HN 2327#undef HN_FUNC_HN_H 2328#undef HN_FUNC_HN_HN_HN 2329#undef HN_FUNC_HN_IN 2330#undef H_FUNC_HN 2331#undef H_FUNC_HN_HN 2332#undef SCALARIZE_HN_FUNC_HN_HN 2333 2334// exports unavailable mathlib functions to compat lib 2335 2336#ifdef RS_COMPATIBILITY_LIB 2337 2338// !!! DANGER !!! 2339// These functions are potentially missing on older Android versions. 2340// Work around the issue by supplying our own variants. 2341// !!! DANGER !!! 2342 2343// The logbl() implementation is taken from the latest bionic/, since 2344// double == long double on Android. 2345extern "C" long double logbl(long double x) { return logb(x); } 2346 2347// __aeabi_idiv0 is a missing function in libcompiler_rt.so, so we just 2348// pick the simplest implementation based on the ARM EABI doc. 2349extern "C" int __aeabi_idiv0(int v) { return v; } 2350 2351#endif // compatibility lib 2352